talent-search-skill

talent-search-skill

An Agent Skill for end-to-end AI / infra / quant talent sourcing — input a candidate persona or a JD, get back a cross-verified, confidence-tiered Excel candidate list plus a full search log. Also covers GitHub contributor mining, target-company recon and roster-page verification.

Works with Claude Code, WorkBuddy, and any AI Agent host that follows the SKILL.md spec.

Why use it

1. Four-channel parallel search — academia (papers / labs / competitions), GitHub (open-source contributors), company (team pages / tech blogs), community (public pages on Zhihu, LinkedIn, Maimai). All four channels fire at once.
2. Cross-verification discipline — every candidate goes through at least one round of primary-source verification (single-pass, mandatory). For critical roles you can switch on five-pass verification — five independent dimensions (academia, GitHub, professional platforms, company primary sources, personal homepage) corroborate each other; conflicts defer to primary sources.
3. Two independent sources per attribution — a name that merely looks right is not evidence. Every identity attribution (this person = this GitHub account / works at this company / authored this paper) needs two independent primary sources; a single path is output as "unconfirmed". Homonyms are the dominant error source in sourcing, so the skill is tuned to miss rather than misfire.
4. Confidence tiers + A/B/C delivery — instead of padding to a headcount, output is tiered: A all hard constraints primary-sourced, B identity verified but graduation year inferred (labelled), C rejected with reason. Each row can carry a confidence cell (high/medium/low, colour-coded) plus the two evidence URLs, so the hiring manager judges trust themselves.
5. Better empty than fabricated — anything not backed by primary sources is marked "to be verified / not obtained". No fabrication. Every key fact is traceable to a source URL in the search log.
6. Zero-dependency Excel output — built-in gen_excel.mjs (Node ≥ 18, no third-party packages) writes a fixed 9-column sheet: name / education / internships-and-jobs / personal homepage / GitHub / first-author papers / university-lab experience / match highlights / notes (CV and contact). Add --confidence to append confidence / delivery tier / evidence columns.
7. A candidate library that grows with you — every round's results land in talent-research/_db/candidates.jsonl. The next round in the same direction queries the library first for incremental updates: anything verified within 90 days is reused, anything rejected is skipped. Paired with db_upsert.mjs (zero-dependency upsert / search / stats commands).
8. Self-improving experience — every round ends with an automatic retrospective. Source effectiveness, signal effectiveness, rejection patterns, network pitfalls, process improvements and tooling notes get appended to monthly files in talent-research/_lessons/. The next round reads the last two months of lessons and tunes its strategy.
9. Evaluation baseline (brake on evolution) — built-in eval_recall.mjs scores every revision against a golden set: recall threshold 0.8; school and current-role field accuracy must not regress. A revision is only allowed to merge if the gate holds — so "self-improvement" can't drift into "self-destruction".
10. Proactive radar (scheduled monitoring) — radar_scan.mjs runs weekly, scanning target companies' GitHub org member changes plus OpenAlex new papers (with first-author and institution). Incremental signals auto-compose a report. From "search when asked" to "proactive talent radar". Ships with a built-in lookup table for 16 China-side LLM-team orgs (Qwen / DeepSeek / Zhipu / Kimi / ModelScope / Tongyi / Tencent ARC & Hunyuan / ByteDance Seed / PaddlePaddle / Noah's Ark Lab / MindSpore / Meituan / JD / RedNote / Xiaohongshu).
11. School tiering built in — C9 / Hong Kong / overseas school tier table and well-known-lab criteria are wired in, so the skill can shortlist by school bar without you re-pasting the list.
12. Field-tested playbooks, not generic advice — references/ ships the hard-won parts: an anonymous/rate-limited GitHub SOP (HTML endpoints over API, email↔account mapping, bot blacklist), a repo × company map per direction, a JD→signal translation recipe with boolean queries and phone-screen questions, company recon, roster-page verification, and site/network pitfalls (including why you must dispatch sub-agents two at a time).

Workflow

Persona / JD
   │
   ▼
Step 0  Parse persona (or translate JD must-haves into strong signals)
        → declare A/B/C delivery tiers → query candidate library (incremental)
        → read last two months of lessons
   │
   ▼
Step 1  Four-channel parallel search  ──→  8–15 candidate lead cards
   │        (max 2 sub-agents at a time; main thread is the fallback)
   │
   ▼
Step 2  Dedupe + cross-verify → write back to candidate library
   │        ├─ Two independent primary sources per attribution (hard rule)
   │        ├─ Single-pass (mandatory): primary-source check on school / current role / contact
   │        ├─ Five-pass (optional): academia / GitHub / professional platform / company / personal homepage
   │        └─ Confidence tier + A/B/C
   │
   ▼
Step 3  Produce candidates-YYYYMMDD.xlsx + search-log.md
   │
   ▼
Step 4  Retrospective  ──→  lessons appended to _lessons/, next round reuses them

Install

Claude Code (user-level, available globally):

git clone https://github.com/cw20051219-commits/talent-search-skill ~/.claude/skills/talent-search

Or project-level (current project only):

git clone https://github.com/cw20051219-commits/talent-search-skill .claude/skills/talent-search

WorkBuddy:

git clone https://github.com/cw20051219-commits/talent-search-skill ~/.workbuddy/skills/talent-search

Other Agent hosts — copy this folder to the host's skills directory. The skill format is a generic SKILL.md + frontmatter spec.

Requirement: Node.js ≥ 18 (only used for the Excel generator, the candidate library and the GitHub probe; zero third-party dependencies).

Usage

After install, just talk to your agent, for example:

Find me 20 candidates for LLM pre-training data roles: PhD 2021–2024, C9 or overseas schools, currently based in China.

Or hand it a benchmark persona / JD:

Here is our benchmark persona (in the examples/persona.example.md format). Find people like this.

Other things it handles out of the box:

Mine the core contributors of verl-project/verl and tell me who owns the training loop.

Recon this company: who are the engineers behind their training stack?

Here's a university HPC team roster page — verify each member's GitHub and current status.

The agent drops into the skill flow automatically. You can also craft inputs by hand using the persona template and candidates.example.json under examples/.

GitHub deep-dive probe

scripts/gh_probe.mjs is a zero-dependency probe for the GitHub channel, built for anonymous / rate-limited networks — it prefers HTML endpoints over the 60-req/hour API:

node scripts/gh_probe.mjs check verl-project/verl microsoft/DeepSpeed   # canonical names; exposes renames
node scripts/gh_probe.mjs contributors <owner/repo> --limit 30          # contributor board (retries 202, flags bots)
node scripts/gh_probe.mjs profile <login> [...]                         # hovercard first, API fallback
node scripts/gh_probe.mjs emailmap <owner/repo> <login>                 # commit-email ↔ account mapping (hardest evidence)
node scripts/gh_probe.mjs codeowners <owner/repo>                       # .github/CODEOWNERS = official module ownership
node scripts/gh_probe.mjs members <org>                                 # org public members (no API quota)
node scripts/gh_probe.mjs repos <org>                                   # org public repos
node scripts/gh_probe.mjs rate                                          # quota self-check
node scripts/gh_probe.mjs selftest                                      # offline self-test (bot blacklist + parsers)

Add --json for machine-readable output. GITHUB_TOKEN is optional (raises the ceiling from 60 to 5000 req/hour; read from env only, never written to disk).

Repo structure

talent-search-skill/
├── SKILL.md                    # Main directive (English): 5-step flow (query / write-back / retrospective)
│                               #   + two-source rule + confidence tiers + output spec
├── SKILL.zh-CN.md              # Same directive in Chinese
├── references/
│   ├── source-playbook.md      # Four-channel search strategy and contact-collection rules
│   ├── school-list.md          # C9 / HK / overseas school tier table and well-known-lab criteria
│   ├── github-playbook.md      # GitHub deep-dive SOP for anonymous/rate-limited networks + pitfalls
│   ├── repo-map.md             # Repo × company map per direction (+ recipe to rebuild for a new direction)
│   ├── sourcing-plan.md        # JD → signal translation, boolean queries, screening rubric, phone screen
│   ├── company-recon.md        # Target-company recon: GitHub org / paper emails / registration / press
│   ├── roster-verify.md        # Official roster page → verified candidate table (confidence tiers)
│   ├── field-notes.md          # Sub-agent limits, output-pollution discipline, site tactics, network env
│   └── lessons-template.md     # Retrospective entry format (six entry types)
├── scripts/
│   ├── gen_excel.mjs           # Zero-dependency xlsx writer (9 columns; --confidence adds 3 more)
│   ├── gh_probe.mjs            # Zero-dependency GitHub probe (check/contributors/profile/emailmap/…)
│   ├── db_upsert.mjs           # Zero-dependency candidate library (upsert / search / stats)
│   ├── eval_recall.mjs         # Zero-dependency golden-set evaluation (recall + field-accuracy gate)
│   └── radar_scan.mjs          # Zero-dependency proactive radar (GitHub org members + OpenAlex paper delta)
├── assets/
│   └── demo.gif                # Workflow demo (fictional data only)
└── examples/
    ├── persona.example.md        # Three-section persona template (fictional people)
    ├── candidates.example.json   # Excel input data structure (incl. confidence / tier / sources)
    ├── db-candidates.example.jsonl  # Candidate library schema (fictional people)
    ├── golden.example.jsonl      # Golden-set format (fictional people)
    └── radar-config.example.json  # Proactive-radar config

Compliance

• This skill only collects publicly available information (GitHub profile emails, personal-homepage emails, paper corresponding-author emails, public LinkedIn pages). No paid databases, no closed-group scraping.
• The candidate data it produces is personal information. You are responsible for using it lawfully (e.g. legitimate recruiting) and for complying with the personal-data laws of your jurisdiction (e.g. PIPL, GDPR, etc.).
• No candidate personal data belongs in the repo. references/ documents method and public organisation assets only — never names, emails or candidate lists. The candidate library, lessons and eval sets are .gitignored.
• The user of this tool bears all compliance responsibility.

License

MIT